Thanks for the amazing implementation. Some questions regarding the implementation details of the model_gcn.py.
1.line 27, why set global feature link adj[-1,:-1]=0? Why not keep adj symmetrical as the rest of the matrix?
2.line 89, why use the distance gap, rather than the distance between part feature and global feature directly as the paper described? In the implementation, it seems the distance gap is essentially the distances between part features without abs?
3.line 102, why *2?
4.line 123, the second graph adgcn2 still uses the original adj as input rather than the learned adj from the previous gcn?
It would be appreciated if you can shed some insights on these questions. Thanks!
Thanks for the amazing implementation. Some questions regarding the implementation details of the model_gcn.py.
1.line 27, why set global feature link adj[-1,:-1]=0? Why not keep adj symmetrical as the rest of the matrix?
2.line 89, why use the distance gap, rather than the distance between part feature and global feature directly as the paper described? In the implementation, it seems the distance gap is essentially the distances between part features without abs?
3.line 102, why *2?
4.line 123, the second graph adgcn2 still uses the original adj as input rather than the learned adj from the previous gcn?
It would be appreciated if you can shed some insights on these questions. Thanks!